A dynamic quasi-newton method for model independent visual servoing

Abstract

This thesis develops and demonstrates a dynamic quasi-Newton method for model independent, vision-guided robotic tracking control. This method does not require precisely calibrated kinematic and camera models. Robotic control is achieved using a method that minimizes a nonlinear objective function by taking quasi-Newton steps and estimating the composite Jacobian at each step. The composite Jacobian combines the image and robot Jacobians and relates differential changes in the robot joint angles to differential changes in the image plane. Two types of Jacobian estimation schemes are investigated, a dynamic Broyden's method and a recursive least squares algorithm. The recursive least squares estimation scheme is found to be more stable in the presence of noise. This method can be applied to vision-guided robotic control by comparing points on the robot with goal points. If the goal points are moving, the objective function must be considered as a function of time. Previous work uses Broyden's method to servo the robot to a static point. It is shown that Broyden's method is not capable of tracking a moving goal. A more appropriate dynamic quasi-Newton method is derived with recursive least squares Jacobian estimation for moving target tracking. Simulation and experimental results demonstrate the dynamic quasi-Newton method for model vision-guided robotic tracking control.